Building Geospatial Foundation Models

Department of Geography

Fall 2025

Advancing environmental monitoring through AI

Course Description

This project-driven seminar teaches students to build geospatial foundation models (GFMs) from scratch. Students implement every layer of the pipeline—from data pipelines and tokenization through attention mechanisms, full architectures, pretraining, evaluation, and deployment—culminating in a working end-to-end GFM tailored to a chosen geospatial application.

By the end of the course, students will be able to:

• Design and implement geospatial data pipelines for multi-spectral, spatial, and temporal data
• Build attention mechanisms and assemble transformer-based architectures for geospatial inputs
• Pretrain using masked autoencoding and evaluate learned representations
• Fine-tune models for specific Earth observation tasks
• Deploy models via APIs and interactive interfaces with honest performance analysis

Getting Started with the UCSB AI Sandbox

Here are detailed instructions for setting up the class environment on the UCSB AI Sandbox, including foundation model installation and GPU optimization. This should all be taken care of for the class, but could be helpful if you are interested in deploying our class infrastructure on a different server or a local machine.

Course Structure: 3 Stages, 10 Weeks

flowchart TD
    subgraph Stage1 ["🏗️ Stage 1: Build GFM Architecture"]
        direction LR
        W1["📊<br/>Week 1<br/>Data Foundations<br/>Pipelines & Tokenization"] --> W2["🧠<br/>Week 2<br/>Attention Mechanisms<br/>Spatial-Temporal Focus"]
        W2 --> W3["🏛️<br/>Week 3<br/>Complete Architecture<br/>Vision Transformer"]
    end
    
    subgraph Stage2 ["🚀 Stage 2: Train Foundation Model"]
        direction LR
        W4["🎭<br/>Week 4<br/>Pretraining<br/>Masked Autoencoder"] --> W5["⚡<br/>Week 5<br/>Training Optimization<br/>Stability & Efficiency"]
        W5 --> W6["📈<br/>Week 6<br/>Evaluation & Analysis<br/>Embeddings & Probing"]
        W6 --> W7["🔗<br/>Week 7<br/>Model Integration<br/>Prithvi, SatMAE"]
    end
    
    subgraph Stage3 ["🎯 Stage 3: Apply & Deploy"]
        direction LR
        W8["🎯<br/>Week 8<br/>Fine-tuning<br/>Task-Specific Training"] --> W9["🚀<br/>Week 9<br/>Deployment<br/>APIs & Interfaces"]
        W9 --> W10["🎤<br/>Week 10<br/>Presentations<br/>Project Synthesis"]
    end
    
    Stage1 --> Stage2
    Stage2 --> Stage3
    
    style Stage1 fill:#e3f2fd
    style Stage2 fill:#fff3e0  
    style Stage3 fill:#e8f5e8
    style W1 fill:#bbdefb
    style W4 fill:#ffe0b2
    style W8 fill:#c8e6c8

🏗️ Stage 1: Build GFM Architecture (Weeks 1-3)

• Week 1: Geospatial Data Foundations (data pipelines, tokenization, loaders)
• Week 2: Spatial-Temporal Attention Mechanisms (from-scratch implementation)
• Week 3: Complete GFM Architecture (Vision Transformer for geospatial)

🚀 Stage 2: Train a Foundation Model (Weeks 4-7)

• Week 4: Pretraining Implementation (masked autoencoder)
• Week 5: Training Loop Optimization (stability, efficiency, mixed precision)
• Week 6: Model Evaluation & Analysis (embeddings, probing, reconstructions)
• Week 7: Integration with Existing Models (Prithvi, SatMAE)

🎯 Stage 3: Apply & Deploy (Weeks 8-10)

• Week 8: Task-Specific Fine-tuning (efficient strategies, few-shot)
• Week 9: Model Implementation & Deployment (APIs, UI, benchmarking)
• Week 10: Project Presentations & Synthesis

Course Sessions

• Weekly sessions: see navbar → 💻 weekly sessions

Teaching Team

Instructor

Kelly Caylor
Email: caylor@ucsb.edu
Learn more: Bren profile

TA

Anna Boser
Email: anaboser@ucsb.edu
Learn more: Bren profile